Tensorboard Scaling for Collaboration

Hi,

We're evaluating some practical improvements we could make to tensorboard.

In particular, enabling it to scale to more data by loading less data at once. Mainly I think this means the collection of runs that's loaded into the application front-end should be something more like an online search, as opposed to what it seems to do now, which is to load everything all at once, and then poll continuously for more runs.

I'm also interested in making sure that the back-end is secure. This will probably mean removing features of the backend until I have as few as possible, and auditing the remaining ones. It might be ideal if the back-end could be entirely removed and replaced with static asset hosting, although that might hurt some of my future ambitions for this project.

I'm also interested in what it would take to make this a public-facing application. This includes things like distributing all static assets without Werkzeug, and disabling features such as Nan/Inf debugging.

Any guidance or discussion along these lines would be greatly appreciated!

I'd also like to know if there is a mailing list, IRC channel, slack or similar where I can learn more about what's going on in tensorboard development, see others' discussion about tensorboard, etc.

I'm also very interested in the file formats used by tensorboard.

Thanks!

No specific file, test, or entry point is named. Start by mapping the front-end run-loading and polling behavior, the back-end features served through Werkzeug, and the static assets; then define separate, reviewable scopes for scaling, security, and public hosting before implementing anything.